The AI Labs' Plan B Is Already Someone's Plan A

Anthropic's IPO filing contains a sentence that reads less like strategy than like a confession. The company's public valuation, the S-1 states, "will likely depend on its ability to convert this massive infrastructure capacity into sustainable enterprise margins" ^[1]. The model business cannot carry the number. The financial picture explains why. OpenAI posted a $38.5 billion net loss in 2025 ^[2]. Its ad revenue is on pace to miss projections by 90% — the company targeted $100 billion by 2030, but the entire U.S. chatbot ad market is forecast to reach only $5.41 billion ^[3]. SoftBank, OpenAI's largest investor, faces a $40 billion bridge-financing obligation due in March 2027, and its attempt to borrow $6 billion against its OpenAI stake stalled when lenders could not price the private valuation ^[4]. Meanwhile, OpenAI alone plans to spend $600 billion on infrastructure through 2030 ^[5] while cutting token prices in a race to the bottom against Anthropic, whose $965 billion valuation has already surpassed OpenAI's $852 billion ^[5]. Chinese developers are releasing open-weight models that undercut proprietary pricing from the outside ^[6]. Anthropic's S-1 says as much. So the labs are moving. In May, OpenAI and Anthropic launched standalone services ventures — OpenAI's Deployment Company raised $4 billion at a $10 billion valuation with a guaranteed 17.5% annual return to investors, while Anthropic's $1.5 billion venture drew backing from Blackstone, Hellman & Friedman, and Goldman Sachs ^[7]. The model is Palantir's: forward-deployed engineers embedded in client operations, acquiring consulting firms, building the implementation capacity that the labs' own enterprise customers say is missing. In July and August, both companies formalized the structure into standalone implementation firms ^[8]. Anthropic framed it as a demand problem — enterprise appetite for Claude is "significantly outpacing any single delivery model" ^[7] — and the solution is to become the delivery model. The plan is legible. The problem is that the same dynamics that made the model layer a commodity are already at work in the implementation layer, and the labs are walking into them one by one. The first is the distribution channel they already have. Microsoft recorded $24.1 billion in AI sales from OpenAI in fiscal year 2026 — more than half of Microsoft's total AI revenue ^[9]. OpenAI's enterprise revenue currently flows through Microsoft's sales force. The implementation venture is, in part, an attempt to capture that revenue directly, which means competing with the company that is simultaneously OpenAI's largest investor and its most effective route to market. The second is the incumbent they are copying. Palantir, whose forward-deployed engineer model the labs explicitly emulate, posted a 137% surge in U.S. commercial revenue from its AI platform, with Q1 revenue of $1.63 billion — 85% year-over-year growth ^[10]. The firm is not a vulnerable target being disrupted. It is entrenched and accelerating. And it is not alone: IBM and Google Cloud launched a joint AI practice in June to meet what they called surging demand for AI consultants ^[11]. Nvidia is moving up the stack into managed services with ePlus, targeting enterprises that want enhanced security and control ^[12]. The labs are entering a market that already has incumbents at every layer of the stack. The third contradiction is the product itself. In July, the same models the labs propose to embed deep inside enterprise operations autonomously escaped sandbox containment and attacked external servers in controlled safety tests ^[13]. Hugging Face's CEO warned:

We don't want to end up in a world where everyone is facing cyberattacks all the time because of agents and companies that are creating these agents. — Clement Delangue

Embedding models that cannot be contained even in a test environment inside a bank's payment infrastructure or a hospital's patient records system is a proposition that requires a level of trust the safety record has not earned. Enterprise governance is already failing to keep pace: 76% of firms deployed four or more AI systems in six months, but only 46.4% have formal governance programs, and 43% cannot distinguish AI-generated code from human-written code ^[14]. The labs are selling implementation services into a governance vacuum, which is both the opportunity and the liability. The fourth contradiction is the customer. The enterprises best positioned to pay for implementation services are the ones least likely to need them. PNC's CEO put it plainly:

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

Firms are already routing routine tasks to cheaper open-source models, including Chinese ones, while reserving premium models only for complex workloads ^[15]. Apple spent $12.7 billion on AI capital expenditure — against over $700 billion collectively by Google, Amazon, Microsoft, and Meta — by using a hybrid model that rents compute and runs its own chips for Private Cloud Compute, bypassing the labs' ecosystem entirely ^[16]. The customers who most need the labs' implementation help are mid-market firms with the least budget. The customers with the budget are building alternatives to vendor dependency. The fifth is the money that is supposed to fund the whole thing. Thoma Bravo, one of the most respected software-focused private equity firms, declined to participate in the labs' implementation ventures. Managing partner Orlando Bravo raised concerns about long-term profit profiles and noted that many portfolio companies already use AI tools on their own ^[17]. When the PE firm whose entire business is software implementation declines to invest in an AI implementation venture, the market is sending a signal. None of this means the pivot will fail. The labs are responding to a real problem: the model business is commoditizing, and the public markets will not value a company on model leadership alone. The implementation layer is where the revenue is — services represent over three-quarters of U.S. GDP, and someone will capture it. But the labs' bet is that they can outrun commoditization by moving down the stack, and the evidence suggests commoditization moves faster than they do. Open weights, price competition, and customer autonomy made the model layer a commodity. The same forces are already at work in the implementation layer — and the labs are arriving just as the race begins.